~753B total / 40B active (MoE; official repo table also labels the family 744B-A40B)
1M
MIT
Official open-weight GLM-5.2 flagship release positioned for long-horizon coding and agentic engineering tasks, with both standard and FP8 checkpoints published. Z.AI describes GLM-5.2 as the first GLM release to deliver a 'solid' 1M-token context for practical long-horizon work rather than only nominal long-context support, alongside up to 128K output tokens, multiple reasoning-effort levels, structured output, function calling, MCP integration, and context caching. The official GitHub README says the model introduces IndexShare, which reuses the same sparse-attention indexer across every four sparse layers to reduce per-token FLOPs by 2.9x at 1M context, and also improves the MTP speculative-decoding layer with up to 20 percent longer acceptance length. Official benchmark tables claim GLM-5.2 is the strongest open-source coding model in its comparison set, improving over GLM-5.1 from 62.0 to 81.0 on Terminal-Bench 2.1 and from 58.4 to 62.1 on SWE-bench Pro, while also ranking highly on FrontierSWE, PostTrainBench, SWE-Marathon, MCP-Atlas, and several reasoning benchmarks. Source materials disagree slightly on scale, with Hugging Face listing 753B parameters while the official GitHub download table labels the release family 744B-A40B; both present it as a frontier-scale sparse MoE GLM-5-series checkpoint. The Hugging Face card also emphasizes a pure MIT open-source license with no regional restrictions and local deployment support across SGLang, vLLM, xLLM, Transformers, and KTransformers.